Opening the book…
While the Mori-Zwanzig (M-Z) formalism provides an exact framework for deriving irreversible macroscopic equations from reversible microscopic dynamics, it is often viewed as a formal mathematical identity. In this chapter, we apply the M-Z projection operator technique to the Magalinski model-a system oscillator coupled to a harmonic bath. We demonstrate that the abstract "orthogonal evolution" corresponds precisely to the free dynamics of the bath. This concrete example serves to ground the formal identities of the previous chapter in physical intuition, explicitly deriving the memory kernel and the fluctuation-dissipation relation.
14.1 Introduction: From Formalism to Physics
As developed in the preceding chapter, the Mori-Zwanzig identity rearranges the Liouville equation into a form resembling the Langevin equation. However, the physical nature of the memory kernel \(K(t)\) and the noise \(F(t)\) often remains opaque.
To bridge this gap, we apply the method to the Magalinski model [1]. This system is uniquely instructive because it allows for an exact evaluation of the projection operators, revealing that the "orthogonal space" is not merely a mathematical construct, but represents the physical degrees of freedom of the environment that we choose to "project out."
14.2 The Coupled Oscillator Hamiltonian
We consider a system oscillator (coordinate \(q_{0}\), momentum \(p_{0}\) ) bilinearly coupled to \(N\) bath oscillators \(\left(q_{i}, p_{i}\right)\). The total Hamiltonian is:
\[ H=\frac{p_{0}^{2}}{2 m_{0}}+\frac{1}{2} m_{0} \omega_{0}^{2} q_{0}^{2}+\sum_{i=1}^{N}\left(\frac{p_{i}^{2}}{2 m_{i}}+\frac{1}{2} m_{i} \omega_{i}^{2} q_{i}^{2}\right)+\sum_{i=1}^{N} c_{i} q_{0} q_{i}, \tag{14.1} \label{14.1} \]
where \(c_{i}\) are the coupling constants. Recalling the Liouville operator \(\mathcal{L} A=\{A, H\}\), we decompose it into system, bath, and interaction components:
\[ \begin{align*} \mathcal{L}_{0} & =\frac{p_{0}}{m_{0}} \frac{\partial}{\partial q_{0}}-m_{0} \omega_{0}^{2} q_{0} \frac{\partial}{\partial p_{0}} \\[4pt] \mathcal{L}_{B} & =\sum_{i=1}^{N}\left(\frac{p_{i}}{m_{i}} \frac{\partial}{\partial q_{i}}-m_{i} \omega_{i}^{2} q_{i} \frac{\partial}{\partial p_{i}}\right), \tag{14.2} \label{14.2}\\[4pt] \mathcal{L}_{\mathrm{int}} & =-\sum_{i=1}^{N} c_{i} q_{i} \frac{\partial}{\partial p_{0}}-\sum_{i=1}^{N} c_{i} q_{0} \frac{\partial}{\partial p_{i}} . \end{align*} \]
14.3 Choosing Relevant Variables
A natural choice of relevant (slow) variables for the system oscillator is the vector
\[ A(t)=\binom{q_{0}(t)}{p_{0}(t)} . \tag{14.3} \label{14.3} \]
The equation of motion for \(A(t)\) is given formally by
\[ \frac{d}{d t} A(t)=\mathcal{L} A(t), \quad A(t)=e^{\mathcal{L} t} A(0) . \tag{14.4} \label{14.4} \]
The action of \(\mathcal{L}\) on \(A\), using the above Hamiltonian, Equation \ref{14.20} is
\[ \mathcal{L} A=\binom{\mathcal{L} q_{0}}{\mathcal{L} p_{0}}=\binom{p_{0} / m_{0}}{-m_{0} \omega_{0}^{2} q_{0}-\sum_{j} c_{j} q_{j}} . \tag{14.5} \label{14.5} \]
(Repeat: Some authors include a factor \(i\) by convention common in projection-operator literature; here we work without it.)
\[ \mathcal{P} B=(B, A)(A, A)^{-1} A=\sum_{j, k=0}^{1}\left\langle A_{j} B\right\rangle_{e q}\left(\mathbb{C}^{-1}\right)_{j k} A_{k}, \tag{14.6} \label{14.6} \]
where \(\mathbb{C}_{j k}=\left\langle A_{j}(0) A_{k}(0)\right\rangle_{e q}\) is the correlation matrix.
With the identification of the relevant vector \(A\) is a 2-component vector, Equation \ref{14.3},
\[ \mathbb{C}=\left(\begin{array}{cc} \left\langle q_{0}^{2}\right\rangle_{e q} & \left\langle q_{0} p_{0}\right\rangle_{e q} \\[4pt] \left\langle p_{0} q_{0}\right\rangle_{e q} & \left\langle p_{0}^{2}\right\rangle_{e q} \end{array}\right) \tag{14.7} \label{14.7} \]
The equipartition theorem [4], remarkably, is independent of the precise form of the function \(H\). An elegant proof is given in the book by Pathria and Beale [4], page 62. Using the equipartition theorem we have,
\[ \mathbb{C}=\left(\begin{array}{cc} \left\langle q_{0}^{2}\right\rangle_{e q} & 0 \\[4pt] 0 & \left\langle p_{0}^{2}\right\rangle_{e q} \end{array}\right)=\left(\begin{array}{cc} \frac{k_{B} T}{m_{0} \omega_{0}^{2}} & 0 \\[4pt] 0 & m_{0} k_{B} T \end{array}\right) . \tag{14.8} \label{14.8} \]
14.3.1 Frequency Matrix:The Systematic Term
Recall that the frequency matrix \(\Omega\) is defined as
\[ \Omega=\langle\mathcal{L} A, A\rangle \mathbb{C}^{-1} \tag{14.9} \label{14.9} \]
First consider the first factor on the right hand side, i. e., the matrix \(\langle\mathcal{L} A, A\rangle\)
\[ \left(\begin{array}{ll} \left\langle q_{0} \mathcal{L} q_{0}\right\rangle_{e q} & \left\langle p_{0} \mathcal{L} q_{0}\right\rangle_{e q} \tag{14.10} \label{14.10}\\[4pt] \left\langle q_{0} \mathcal{L} p_{0}\right\rangle_{e q} & \left\langle p_{0} \mathcal{L} p_{0}\right\rangle_{e q} . \end{array}\right) \]
Now use Eq. 14.5 and note that the diagonal terms involve cross-terms between \(\sim q_{0} p_{0}\), i. e., the 11 term involves \(\left\langle q_{0} \mathcal{L} q_{0}\right\rangle_{e q} \sim\left\langle q_{0} p_{0}\right\rangle_{e q}\) and the 22 term involves \(\left\langle p_{0} \mathcal{L} p_{0}\right\rangle_{e q} \sim\left\langle p_{0} q_{0}\right\rangle_{e q}\) and, therefore, they Vanish. Similarly, the off-diagonal term12 is \(\left\langle p_{0} \mathcal{L} q_{0}\right\rangle_{e q}=\left\langle p_{0}^{2} /\left(m_{0}\right)\right\rangle_{e q}=k_{B} T\) and the off-diagonal term 21 is \(\left.\left\langle q_{0} \mathcal{L} p_{0}\right\rangle_{e q}=\left\langle q_{0}\left(-m_{0} \omega_{0}^{2} q_{0}\right)\right)\right\rangle_{e q}=-k_{B} T\)
\[ \left(\begin{array}{ll} \left\langle q_{0} \mathcal{L} q_{0}\right\rangle_{e q} & \left\langle p_{0} \mathcal{L} q_{0}\right\rangle_{e q} \\[4pt] \left\langle q_{0} \mathcal{L} p_{0}\right\rangle_{e q} & \left\langle p_{0} \mathcal{L} p_{0}\right\rangle_{e q} \end{array}\right)=\left(\begin{array}{cc} 0 & k_{B} T \\[4pt] -k_{B} T & 0 \end{array}\right) . \tag{14.11} \label{14.11} \]
Using \(\mathbb{C}^{-1}=\operatorname{diag}\left(\frac{m_{0} \omega_{0}^{2}}{k_{B} T}, \frac{1}{m_{0} k_{B} T}\right)\) one obtains the expected structure for \(\boldsymbol{\Omega}\)
\[ \Omega=\left(\begin{array}{ll} \left\langle q_{0} \mathcal{L} q_{0}\right\rangle_{e q} & \left\langle p_{0} \mathcal{L} q_{0}\right\rangle_{e q} \\[4pt] \left\langle q_{0} \mathcal{L} p_{0}\right\rangle_{e q} & \left\langle p_{0} \mathcal{L} p_{0}\right\rangle_{e q} \end{array}\right) \mathbb{C}^{-1}=\left(\begin{array}{cc} 0 & \frac{1}{m_{0}} \\[4pt] -m_{0} \omega_{0}^{2} & 0 \end{array}\right) \tag{14.12} \label{14.12} \]
14.3.2 Noise Vector
We have seen above, Equation \ref{13.15}, that the noise vector is given by \(F(t)=e^{\mathcal{Q} \mathcal{L} t} \mathcal{Q} \mathcal{L} A(0)\). But from Equation \ref{14.5} for \(\mathcal{L} A\) upon further projection by the \(\mathcal{Q}\) yields:
\[ \mathcal{Q} \mathcal{L} A=\mathcal{Q}\binom{p_{0} / m_{0}}{-m_{0} \omega_{0}^{2} q_{0}-\sum_{j} c_{j} q_{j}}=\binom{0}{-\sum_{j} c_{j} q_{j}}=\binom{0}{f(0)} . \tag{14.13} \label{14.13} \]
Since \(\mathcal{L} q_{0}=p_{0} / m_{0}\) lies in the span of the relevant variables, \(\mathcal{Q} \mathcal{L} q_{0}=0\). For the momentum component,
\[ \mathcal{Q} \mathcal{L} p_{0}=-\sum_{i=1}^{N} c_{i} q_{i} \equiv f(0) . \nonumber \]
14.4 On the projected subspace \(\mathcal{Q}\) the bath evolves freely
For the linear coupling \(H_{\text {int }}=\sum_{i=1}^{N} c_{i} q_{0} q_{i}\), we show
\[ \mathcal{Q} \mathcal{L} f(0)=\mathcal{L}_{B} f(0), \nonumber \]
where \(f(0)=-\sum_{j} c_{j} q_{j}(0)\) and \(\mathcal{L}_{B}\) is the bath Liouville operator generated by \(H_{B}\). Consequently,
\[ F_{2}(t)=e^{\mathcal{Q} \mathcal{L} t} f(0)=e^{\mathcal{L}_{B} t} f(0), \nonumber \]
i.e. the orthogonal evolution of \(f(0)\) is identical to the free bath evolution.
This result is the 'Rosetta Stone' of the derivation, translating the abstract orthogonal dynamics of Mori-Zwanzig into the familiar language of free Newtonian oscillations.
14.4.1 Algebraic proof
First compute
\[ \mathcal{L} f(0)=-\sum_{j} c_{j}\left\{q_{j}, H\right\}=-\sum_{j} c_{j} \frac{p_{j}}{m_{j}} . \nonumber \]
Since the right hand side involves only the vectors in the \(B\) space, the further projection by \(\mathcal{Q}\) does not affect it, i.e., \(\mathcal{Q} \mathcal{L} f(0)=-\sum_{j} c_{j} \frac{p_{j}}{m_{j}}\). But this equals \(\mathcal{L}_{B} f(0)\), since from the above definition of \(\mathcal{L}_{B}\), we have, \(\mathcal{L}_{B} q_{j}=p_{j} / m_{j}\). Therefore
\[ \mathcal{Q} \mathcal{L} f(0)=\mathcal{L}_{B} f(0), \nonumber \]
and exponentiation yields \(e^{\mathcal{Q} \mathcal{L} t} f(0)=e^{\mathcal{L}_{B} t} f(0)\).
14.5 Explicit expression for \(F_{2}(t)=e^{\mathcal{Q} \mathcal{L} t} f(0)\)
For the harmonic bath, the homogeneous solution is given by, \(q_{j}^{\text {hom }}(t)\),
\[ e^{\mathcal{L}_{B} t} q_{j}(0)=q_{j}(0) \cos \left(\omega_{j} t\right)+\frac{p_{j}(0)}{m_{j} \omega_{j}} \sin \left(\omega_{j} t\right)=q_{j}^{\mathrm{hom}}(t) . \nonumber \]
Thus,
\[ F_{2}(t)=-\sum_{j=1}^{N} c_{j}\left[q_{j}(0) \cos \left(\omega_{j} t\right)+\frac{p_{j}(0)}{m_{j} \omega_{j}} \sin \left(\omega_{j} t\right)\right] . \tag{14.14} \label{14.14} \]
i.e. the noise vector is
\[ F(t)=\binom{0}{F_{2}(t)}=\binom{0}{e^{\mathcal{Q} \mathcal{L} t} f(0)} . \nonumber \]
From now on we will designate the non-zero component of the above vector as \(F(t)=F_{2}(t)\), i. e. drop the subscript 2.
14.6 Difference between \(f(t)=e^{\mathcal{L} t} f(0)\) and \(F(t)=F_{2}(t)=e^{\mathcal{Q} \mathcal{L} t} f(0)\)
- Full evolution \(f(t)\). Contains both homogeneous bath evolution and driven contributions from the system degrees of freedom because \(\mathcal{L}\) includes interactions.
- Orthogonal evolution \(F(t)=F_{2}(t)\). Evolution under the dynamics constrained to the subspace orthogonal to the relevant variables. For the linear harmonic model it reduces to the free bath evolution of the same initial combination and contains no driven/system-coupling pieces.
We re-emphasize that \(F(t)\) is "random" only in the sense of being orthogonal to the system variables, but its underlying dynamics are still deterministic.
14.7 Memory Kernel
Next evaluate the general formula
\[ K(t)=\langle F(0), F(t)\rangle_{e q} \mathbb{C}^{-1} \tag{14.15} \label{14.15} \]
for the problem of the coupled oscillator. The non-zero component of the memory kernel is
\[ K_{22}(t)=K_{p p}(t)=\frac{1}{m_{0} k_{B} T}\langle F(0) F(t)\rangle_{e q}, \tag{14.16} \label{14.16} \]
and using Equation \ref{14.14} one obtains
\[ \langle F(0) F(t)\rangle_{e q}=\sum_{i=1}^{N} c_{i}^{2}\left\langle q_{i}(0) q_{i}(t)\right\rangle_{e q}=\sum_{i=1}^{N} \frac{c_{i}^{2} k_{B} T}{m_{i} \omega_{i}^{2}} \cos \left(\omega_{i} t\right) . \nonumber \]
Thus, redefining non-zero component of the memory kernel \(K\) as a scalar, for convenience,
\[ K(t)=K_{22}(t)=K_{p p}(t)=\frac{1}{m_{0}} \sum_{i=1}^{N} \frac{c_{i}^{2}}{m_{i} \omega_{i}^{2}} \cos \left(\omega_{i} t\right) . \tag{14.17} \label{14.17} \]
14.8 Complete Generalized Langevin Equation[4]
By combing the results above for the system oscillator, the frequency matrix, Equation \ref{14.12}, the Kernel Equation \ref{14.17} and the noise Equation \ref{14.14}, with the M-Z formalism Equation (B.10) gives us
\[ \begin{align*} \frac{d q_{0}}{d t} & =\frac{p_{0}}{m_{0}}, \tag{14.18} \label{14.18}\\[4pt] \frac{d p_{0}}{d t} & =-m_{0} \omega_{0}^{2} q_{0}-\int_{0}^{t} K(t-s) \frac{p_{0}(s)}{m_{0}} d s+F(t), \tag{14.19} \end{align*} \]
with the fluctuation-dissipation relation[4]
\[ \langle F(t) F(s)\rangle_{e q}=k_{B} T m_{0} K(|t-s|) . \nonumber \]
(Note: \(\mathbb{C}_{p p}^{-1}=\frac{1}{m_{0} k_{B} T}\) was used in \(K_{22}=K_{p p}=K\) which leads to a factor of \(m_{0}\) here compared to some other conventions.)
14.9 Continuum limit
With the spectral density
\[ J(\omega)=\frac{\pi}{2} \sum_{i=1}^{N} \frac{c_{i}^{2}}{m_{i} \omega_{i}^{2}} \delta\left(\omega-\omega_{i}\right), \nonumber \]
the continuum limit gives
\[ K(t)=K_{22}(t)=\frac{2}{m_{0} \pi} \int_{0}^{\infty} d \omega J(\omega) \cos (\omega t) \nonumber \]
14.9.1 Ohmic damping
For an Ohmic spectral density \(J(\omega)=m_{0} \gamma \omega\) (with appropriate high-frequency cutoff), one finds in the Markov limit
\[ K(t) \approx 2 \gamma \delta(t), \nonumber \]
leading to the classical damped oscillator
\[ \ddot{q}_{0}+2 \gamma \dot{q}_{0}+\omega_{0}^{2} q_{0}=\frac{F(t)}{m_{0}} . \nonumber \]
This example demonstrates how memory kernels reduce to simple friction in the Markovian limit.
14.10 Relation to the generalized Langevin equation of Magalinski[1]
The final results agree with the solution obtained by Magalinski [1] by direct integration (Green's function) solution of coupled harmonic oscillators. We leave Magalinski's [1] derivation as an exercise below. In particular, the noise time evolution is governed by free bath oscillators, exactly the same way.
14.11 Numerical Illustration of "Deterministic" Noise
Even with a finite \(N\), the superposition of sines and cosines in Equation \ref{14.14} produces a "noisy" trajectory. Figure 14.1 illustrates \(F(t)\) for \(N=10\), and its Fourier spectrum is shown in Figure 14.2. Though the underlying equations are deterministic, the ignorance of initial bath conditions \(\rho(\Gamma)\) provides the statistical randomness required for the Langevin description.
14.12 Bilinear Coupling
Here we have assumed a bilinear coupling between the central oscillator and the bath oscillators,
\[ H=\underbrace{\frac{p_{0}^{2}}{2 m_{0}}+\frac{1}{2} m_{0} \omega_{0}^{2} q_{0}^{2}}_{H_{0}}+\underbrace{\sum_{i=1}^{N}\left(\frac{p_{i}^{2}}{2 m_{i}}+\frac{1}{2} m_{i} \omega_{i}^{2} q_{i}^{2}\right)}_{H_{B}}+\underbrace{\sum_{i=1}^{N} c_{i} q_{0} q_{i}}_{H_{\mathrm{int}}} . \tag{14.20} \label{14.20} \]
It is crucial at this stage to alert the reader to a major mathematical specialty of bilinear systems. In general, \(f(0)=\mathcal{Q} \mathcal{L} p_{0}(0)\) is the initial unprojected force generated by the interaction. For linear coupling, \(f(0)\) happens to depend only on the bath coordinates \(\left\{q_{j}\right\}\), containing no primary system coordinates \(q_{0}\) or \(p_{0}\). Because \(f(0)\) resides exclusively in the linear subspace of bath degrees of freedom, applying \(\mathcal{Q} \mathcal{L}\) repeatedly generates operations that stay entirely within the bath subspace.
CHAPTER 14. THE MAGALINSKI MODEL: A PHYSICAL REALIZATION OF MORI-ZWANZIG DYNAM
Figure \(\PageIndex{1}\): Time evolution of the fluctuating force \(F(t)=-\sum_{j=1}^{N} c_{j}\left[q_{j}(0) \cos \left(\omega_{j} t\right)+\right.\) \(\left.\frac{p_{j}(0)}{m_{j} \omega_{j}} \sin \left(\omega_{j} t\right)\right]\) for \(N=10\) bath oscillators with random initial conditions. Despite being generated by fully deterministic equations of motion, the superposition of incommensurate frequencies produces a trajectory that is effectively random.
However, if \(f(0)\) contains higher-order non-linear terms in \(q_{0}\) or cross-terms (such as quadratic/cubic system terms arising from non-linear coupling \(H_{\text {int }}\) or non-linear system potentials), additional terms generated in \(f(0)\) will fail to remain orthogonal to the projection subspace under further applications of \(\mathcal{L}\). Projection operator \(\mathcal{Q}=1-\mathcal{P}\) will then continually strip off non-zero projections back onto \(q_{0}\) and \(p_{0}\), creating an infinite hierarchy of non-linear operator couplings. Thus, the identity \(\mathcal{Q} \mathcal{L} f(0)=\mathcal{L}_{B} f(0)\) is a unique, remarkable closure property of linear bath systems, allowing \(e^{\mathcal{Q} \mathcal{L} t}\) to be replaced by the free bath operator \(e^{\mathcal{L}_{B} t}\).
It is essential to separate two distinct claims. The Mori-Zwanzig identity is an exact operator rearrangement of the Liouville equation: it holds for any Hamiltonian, linear or non-linear, and does not rely on the harmonic bath model at all. What is special to the harmonic bath is not the identity itself but the fact that it can be evaluated in closed form - the exact evaluation \(\mathcal{Q} \mathcal{L} f(0)=\mathcal{L}_{B} f(0)\) used throughout relies crucially on the linear nature of the harmonic interactions between the oscillator and the bath variables. When non-linearities are introduced, either through an anharmonic system potential \(V\left(q_{0}\right)\) or non-linear coupling \(H_{\text {int }}\left(q_{0}, q_{i}\right)\), the identity remains exact but its evaluation becomes intractable in closed form. Students should not read this section as a failure of the Mori-Zwanzig method, but as a statement about which systems admit an explicit solution.
The mechanism is concrete. Suppose the coupling contains a cubic term, \(H_{\text {int }}=c q_{0} q_{j}+\) \(\lambda q_{0}^{2} q_{j}\). Acting with \(\mathcal{Q} \mathcal{L}\) on \(f(0)=q_{j}(0)\) now produces, alongside the linear term \(\propto q_{0}\), a term \(\propto q_{0}^{2}\). Because \(q_{0}^{2}\) does not lie in the linear span of the bath coordinates \(\left\{q_{j}, p_{j}\right\}\) that \(\mathcal{Q}\) projects onto, this term is not annihilated and does not close back into \(\mathcal{L}_{B} f(0)\) : the orthogonal dynamics generates an ever-growing hierarchy of higher powers of \(q_{0}\) under repeated application of \(\mathcal{Q} \mathcal{L}\). Consequently: 1. The orthogonal propagator \(e^{\mathcal{Q} \mathcal{L} t}\) no longer reduces to uncoupled bath dynamics \(\mathcal{L}_{B}\). 2. The fluctuating force \(F(t)\) and memory kernel \(K(t)\) become state-dependent, leading to non-Gaussian noise statistics and non-linear memory structures. 3. Closed-form analytic evaluation of \(K(t)\) is generally impossible, requiring mode-coupling approximations.
Figure \(\PageIndex{2}\): Fourier spectrum of \(F(t)\) (frequencies up to \(\omega=30\) ) for the same \(N=10\) realisation shown in Figure \(\PageIndex{1}\). Each spike corresponds to a bath oscillator frequency \(\omega_{j}\). In the continuum limit \(N \rightarrow \infty\) the discrete spikes merge into the smooth spectral density \(J(\omega)\), recovering the memory kernel \(\gamma(t)\) via \(\gamma(t)=\frac{2}{M \pi} \int_{0}^{\infty} \frac{J(\omega)}{\omega} \cos (\omega t) d \omega\).
Highlighting this limitation helps students appreciate why the harmonic bath model is a unique paradigm where Mori-Zwanzig projection yields fully analytic closed-form solutions, while the identity itself remains available as an exact starting point for perturbative or numerical treatments of the general case.
14.13 Conclusion
The Magalinski model demonstrates that the "memory" in the Generalized Langevin Equation is physically the "echo" of the bath. After the system pushes the bath, and the bath, evolving through its own dynamics, pushes back at a later time. The formal M-Z identity thus finds a concrete, intuitive home in the dynamics of coupled oscillators.
Bibliography
[1] V. I. Magalinski, JETP 36, 1942 (1959).
[2] R. Zwanzig, Nonequilibrium Statistical Mechanics (Oxford, 2001).
[3] R. K. Pathria and P. D. Beale, Statistical Mechanics (Elsevier, 2011).
This page titled 14: The Magalinski Model- A Physical Realization of Mori-Zwanzig Dynamics was last modified on Mon, 02 Nov 2026 06:16:51 GMT and is shared under a CC BY-NC-SA 4.0 license and was authored, remixed, and/or curated by Pabitra N. Sen.
The book could not be loaded just now.